By the end of this lesson you will be able to
- Sort a task into good-fit, needs-supervision, or bad-fit in seconds
- Explain why some limits are structural rather than temporary
- Avoid the two most expensive mis-assignments
The most common expensive mistake with AI is not using it badly. It is pointing it at the wrong task and concluding the technology does not work.
The sorting rule
Ask two questions about the task. Is the input already in front of it? And is there one right answer that someone could check?
| Input supplied? | Checkable right answer? | Verdict |
|---|---|---|
| Yes | No — judgement of quality | Excellent fit. Rewriting, summarising, tone shifts. |
| Yes | Yes | Good fit with a check. Extraction, classification, formatting. |
| No | No | Use as a thinking partner. Brainstorming, first drafts, options. |
| No | Yes | Worst fit. Specific facts recalled from memory. Verify everything. |
Where it is genuinely better than a person
Not "acceptable" — actually better, on the dimensions that matter for the task:
- Never getting bored. The four-hundredth invoice gets the same attention as the first, which is not true of any human.
- Instant format conversion. Notes to email, transcript to summary, spreadsheet to prose, all in seconds.
- Breadth without depth. A working answer across law, marketing, code, and biology in one conversation.
- Producing volume on demand. Twenty subject lines to react to is more useful than a blank page, even if eighteen are discarded.
- Being asked the obvious question with no social cost. People will ask AI what they are embarrassed to ask a colleague.
Where it is structurally bad
These are not gaps that a better model closes. They follow from what the system is, so they are worth understanding rather than waiting out.
Knowing what it does not know
There is no internal signal separating "recalled reliably" from "generated plausibly". It cannot flag its own uncertainty because it does not have access to it.
Anything depending on current reality
Today's price, this week's rule change, whether that company still exists. Unless a tool actively looks it up, the answer comes from a frozen snapshot.
Accountability
Someone has to be answerable for the advice, the number, the decision. That cannot be delegated to a system, which matters most in exactly the regulated fields where automation looks most attractive.
Knowing your unwritten context
That this client is difficult, that the deadline is soft, that the last person who tried this got burned. None of it is written down anywhere the model can see.
AI does not replace the person who knows why the work is being done. It replaces the part of their day spent typing things they already knew.
The two expensive mis-assignments
Over-trusting
- Sending AI output to a client unread
- Using recalled figures in a proposal
- Automating a decision nobody reviews
- Costs: reputation, occasionally worse
Under-using
- Typing the same follow-up for the fiftieth time
- Reading a transcript to find one decision
- Reformatting notes by hand
- Costs: hours a week, invisibly, forever
Over-trusting gets all the press because the failures are dramatic. Under-using is far more common and, across a year, usually more expensive.
Will the structural limits get solved?
Some soften. Retrieval tools largely fix the "current reality" gap by looking things up instead of recalling. Accountability is not a technical problem and will not be solved technically.
What is the best first task to try?
Something you already do repeatedly, where you can immediately tell whether the output is good. Summarising your own meeting notes is close to ideal — you know the right answer, so you can judge quality instantly.
Key takeaways
- Transforming supplied text: excellent. Recalling unsupplied facts: weak.
- Some limits are structural — self-knowledge, current reality, accountability, your unwritten context.
- Over-trusting fails loudly. Under-using fails quietly and usually costs more.
- Start with a repetitive task where you can judge the output instantly.